I built an AI health app with $0/month infrastructure. Here's what the "AI in healthcare" conversation keeps getting wrong.
Every week there's a new headline about AI transforming healthcare. Most of it is about large hospital systems, diagnostic radiology, and drug discovery.
Meanwhile, I spent the last several months building a consumer health app — on-device AI skin analysis, dermatologist booking, personalised routines — and the gap between what the industry talks about and what actually works at the consumer layer is significant.
A few things I learned:
- Cloud inference is the wrong model for consumer health AI
Every "AI health app" I researched bills per inference. For a skincare app, users scan their face daily — that's ~30 API calls per user per month.
At scale, your entire margin disappears before you've collected a dollar.
On-device inference (TFLite in our case) flips this completely:
Zero inference cost
Zero latency
Works offline
For consumer health, where daily usage is the core mechanic — on-device is the only architecture that makes economic sense.
- The data privacy problem solves itself when you don't send data anywhere
HIPAA, GDPR, consent flows, data handling policies — most of this complexity exists because the data leaves the device.
When the AI runs locally:
The scan never leaves the phone
Trust increases significantly
Regulatory burden drops
The privacy story becomes simple.
- $0/month backend is a real product decision, not a hack
Glow Journal (the app I built) runs on:
Google Sheets
Apps Script
Cloudinary
Netlify
Total cost: $0/month.
People assume this won’t scale — and they’re right, eventually.
But “won’t scale past 50K users” is a good problem to have.
The repository pattern is already in place. Migration to Postgres/Firestore is a defined step. Starting free and scaling later is often cheaper than overpaying for unused infrastructure.
- The real moat in consumer health AI isn’t the model — it’s the network
The skin analysis model is table stakes. Anyone can fine-tune a TFLite model.
The defensible asset is the dermatologist network:
Each verified derm brings existing patients
Each clinic onboarding brings hundreds of users
The AI is the hook.
The network is the moat.
I'm actually selling Glow Journal now — looking for someone who wants to launch and grow it rather than build from scratch.
$10K → Code-only
$15K → Live on both stores + Stripe
$25K → Full database migration
But regardless of the sale — the above is simply what I observed building in this space for real.
Curious if anyone else here is building consumer health AI and hitting similar friction points.
What does your inference cost look like?
Are you running on-device or cloud?
Really insightful breakdown of the economics — especially around on-device inference.
From a clinical standpoint though, I think the harder problem in consumer health AI is not cost, but accuracy and risk. A daily-use skin scanner sounds great, but even small diagnostic errors can lead to wrong self-treatment or delayed care.
Do you see this more as a “screening/engagement tool” rather than a diagnostic one? And how tightly is it integrated with dermatologist validation?